On 9/17/21 3:18 PM, Marcel Holtmann wrote:
> Hi Jonas,
>
>>>> The Marvell 88W8897 combined wifi and bluetooth card (pcie+usb version)
>>>> is used in a lot of Microsoft Surface devices, and all those devices
>>>> suffer from very low 2.4GHz wifi connection speeds while bluetooth is
>>>> enabled. The reason for that is that the default passive scanning
>>>> interval for Bluetooth Low Energy devices is quite high on Linux
>>>> (interval of 60 msec and scan window of 30 msec, see le_scan_interval
>>>> and le_scan_window in hci_core.c), and the Marvell chip is known for its
>>>> bad bt+wifi coexisting performance.
>>>>
>>>> So decrease that passive scan interval and make the scan window shorter
>>>> on this particular device to allow for spending more time transmitting
>>>> wifi signals: The new scan interval is 250 msec (0x190 * 0.625 msec) and
>>>> the new scan window is 6.25 msec (0xa * 0.625 msec).
>>>>
>>>> This change has a very large impact on the 2.4GHz wifi speeds and gets
>>>> it up to performance comparable with the Windows driver, which seems to
>>>> apply a similar quirk.
>>>>
>>>> The scan interval and scan window length were tested and found to work
>>>> very well with a bunch of Bluetooth Low Energy devices, including the
>>>> Surface Pen, a Bluetooth Speaker and two modern Bluetooth headphones.
>>>> All devices were discovered immediately after turning them on. Even
>>>> lower values were also tested, but these introduced longer delays until
>>>> devices get discovered.
>>>>
>>>> Signed-off-by: Jonas Dreßler <[email protected]>
>>>> ---
>>>> drivers/bluetooth/btusb.c | 15 +++++++++++++++
>>>> 1 file changed, 15 insertions(+)
>>>>
>>>> diff --git a/drivers/bluetooth/btusb.c b/drivers/bluetooth/btusb.c
>>>> index 60d2fce59a71..05b11179c839 100644
>>>> --- a/drivers/bluetooth/btusb.c
>>>> +++ b/drivers/bluetooth/btusb.c
>>>> @@ -59,6 +59,7 @@ static struct usb_driver btusb_driver;
>>>> #define BTUSB_WIDEBAND_SPEECH 0x400000
>>>> #define BTUSB_VALID_LE_STATES 0x800000
>>>> #define BTUSB_QCA_WCN6855 0x1000000
>>>> +#define BTUSB_LOWER_LESCAN_INTERVAL 0x2000000
>>>> #define BTUSB_INTEL_BROKEN_INITIAL_NCMD 0x4000000
>>>>
>>>> static const struct usb_device_id btusb_table[] = {
>>>> @@ -356,6 +357,7 @@ static const struct usb_device_id blacklist_table[] = {
>>>> { USB_DEVICE(0x1286, 0x2044), .driver_info = BTUSB_MARVELL },
>>>> { USB_DEVICE(0x1286, 0x2046), .driver_info = BTUSB_MARVELL },
>>>> { USB_DEVICE(0x1286, 0x204e), .driver_info = BTUSB_MARVELL },
>>>> + { USB_DEVICE(0x1286, 0x204c), .driver_info = BTUSB_LOWER_LESCAN_INTERVAL },
>>>>
>>>> /* Intel Bluetooth devices */
>>>> { USB_DEVICE(0x8087, 0x0025), .driver_info = BTUSB_INTEL_COMBINED },
>>>> @@ -3813,6 +3815,19 @@ static int btusb_probe(struct usb_interface *intf,
>>>> if (id->driver_info & BTUSB_MARVELL)
>>>> hdev->set_bdaddr = btusb_set_bdaddr_marvell;
>>>>
>>>> + /* The Marvell 88W8897 combined wifi and bluetooth card is known for
>>>> + * very bad bt+wifi coexisting performance.
>>>> + *
>>>> + * Decrease the passive BT Low Energy scan interval a bit
>>>> + * (0x0190 * 0.625 msec = 250 msec) and make the scan window shorter
>>>> + * (0x000a * 0,625 msec = 6.25 msec). This allows for significantly
>>>> + * higher wifi throughput while passively scanning for BT LE devices.
>>>> + */
>>>> + if (id->driver_info & BTUSB_LOWER_LESCAN_INTERVAL) {
>>>> + hdev->le_scan_interval = 0x0190;
>>>> + hdev->le_scan_window = 0x000a;
>>>> + }
>>>> +
>>> you can not do it this way. Modifying hci_dev internals from within the driver is not acceptable.
>>> Regards
>>> Marcel
>>
>>
>> hmm okay, it seems to me that the intention of your commit bef64738e3fb87eabc6fbeededad0c44ea173384 was to allow configuring it on a per controller basis, also btusb changes a bunch of other hci_dev properties? Given that we also have to match by usb-id, I don't think there's another place to do that other than the usb driver, or is there?
>
> you can change most defaults via mgmt commands.
>
> The things the a driver should set in hci_dev is really limited and it only affects its ability to run as a transport driver. It shouldn’t deal with anything that is actually HCI upper layer operation.
>
> Regards
>
> Marcel
>
Hmm, having users of this wifi chip run a script at boot doesn't sound
great to me, or can you point me to somewhere in bluez where I could put
a device-specific quirk for this? IMHO this is a real bug with the
hardware that needs to be quirked somewhere, note that this is not a
minor performance improvement, this bumps wifi 2.4ghz speeds up from 800
KB/s to >5 MB/s.
Other than that, there's also the possibility of changing the
kernel-wide defaults, but as mentioned in my other answer, I don't have
enough hardware laying around to test that at all.
Jonas